I thought it was regression which was at issue here. I'm not convinced cache is so different from all the other aspects of performance we have to deal with in

I thought it was regression which was at issue here. I'm not convinced cache is so different from all the other aspects of performance we have to deal with in HPC systems (insofar as they're isolated), but no matter. At least there's plenty of tool support and literature.